Stable On-Policy Distillation through Adaptive Target Reformulation (2026.findings-acl)
Copied to clipboard
| Challenge: | Knowledge distillation (KD) is widely used for transferring capabilities from proprietary models to efficient open-source counterparts. |
| Approach: | They propose a method that constructs a geometric target distribution in logit space to emphasize agreement between the teacher and the student. |
| Outcome: | Experiments show that the proposed method outperforms supervised fine-tuning and existing on-policy baselines. |
Similar Papers
Fast and Effective On-Policy Distillation from Reasoning Prefixes (2026.findings-acl)
Copied to clipboard
Dongxu Zhang, Zhichao Yang, Sepehr Janghorbani, Jun Han, Andrew Ressler II, Qian Qian, Gregory D Lyng, Sanjit Singh Batra, Robert E. Tillman
| Challenge: | On-policy distillation (OPD) requires expensive on-the-fly sampling of the student policy during training, which substantially increases training cost. |
| Approach: | They propose to use on-policy distillation to sample trajectories from student model . they propose to terminate the sampling early during distillation . |
| Outcome: | The proposed method matches the performance of full OPD in long reasoning outputs while reducing training FLOP by 2x–40x. |
Revisiting Knowledge Distillation for Autoregressive Language Models (2024.acl-long)
Copied to clipboard
| Challenge: | Autoregressive language models (LMs) are expensive and memory intensive, preventing the development of industrial applications. |
| Approach: | They propose an adaptive teaching approach to improve the KD of autoregressive language models by distilling knowledge into a small student model. |
| Outcome: | The proposed method can achieve consistent and significant performance gains across all model types and sizes. |
Calibrated Progressive Distillation: Co-Designing Curriculum and Target Mixing for Knowledge Distillation of Large Language Models (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing methods for knowledge distillation address the teacher–student capacity gap by mixing teacher and student distributions in the distillation target or using curriculum learning to sequence training from easy to hard examples. |
| Approach: | They propose a white-box KD framework that co-designs curriculum scheduling and target mixing through a unified difficulty-aware principle. |
| Outcome: | The proposed framework outperforms existing methods while reducing training runtime by over 10%. |
Cost-effective Distillation of Large Language Models (2023.findings-acl)
Copied to clipboard
| Challenge: | Existing knowledge distillation methods require pretraining of the teacher on task-specific datasets, which can be costly for large and unstable for small datasets. |
| Approach: | They propose an approach to improve knowledge distillation by a loss-agnostic approach to task and model architecture. |
| Outcome: | The proposed method achieves competitive results across a range of tasks, especially for tasks with smaller datasets. |
ToDi: Token-wise Distillation via Fine-Grained Divergence Control (2025.emnlp-main)
Copied to clipboard
| Challenge: | Large language models (LLMs) offer impressive performance but are impractical for resource-constrained deployment due to high latency and energy consumption. |
| Approach: | They propose a method that adaptively combines FKL and RKL per token using a sigmoid-based weighting function derived from the teacher-student probability log-ratio. |
| Outcome: | The proposed method outperforms baselines using uniform or less granular strategies across instruction-following benchmarks. |
Staged Knowledge Distillation Through Least-to-Most Prompting: Optimizing Teacher Guidance via Difficulty-Aware Training (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Knowledge distillation (KD) enables the compression of large language models (LLMs) conventional methods suffer from training-inference mismatches and suboptimal performance due to expensive student-generated outputs. |
| Approach: | They propose a method that combines a CL strategy and adaptive loss design to reduce training mismatches and suboptimal performance. |
| Outcome: | L2M-KD outperforms existing white-box KD methods on instruction-following tasks . it outperformed existing methods, achieving superior student model performance with reduced overhead . |
ReAugKD: Retrieval-Augmented Knowledge Distillation For Pre-trained Language Models (2023.acl-short)
Copied to clipboard
Jianyi Zhang, Aashiq Muhamed, Aditya Anantharaman, Guoyin Wang, Changyou Chen, Kai Zhong, Qingjun Cui, Yi Xu, Belinda Zeng, Trishul Chilimbi, Yiran Chen
| Challenge: | Knowledge distillation (KD) is an effective compression technique to derive a smaller student model from a larger teacher model by transferring the knowledge embedded in the teacher's network. |
| Approach: | They propose a framework and loss function that preserves the semantic similarities of teacher and student training examples to enable the student to retrieve from the knowledge base effectively. |
| Outcome: | The proposed framework preserves the semantic similarities of teacher and student training examples to achieve state-of-the-art performance on the GLUE benchmark. |
Distillation Traps and Guards: A Calibration Knob for LLM Distillability (2026.acl-long)
Copied to clipboard
| Challenge: | Knowledge distillation (KD) transfers capabilities from large language models (LLMs) to smaller students, yet it can fail unpredictably and also underpins model leakage risks. |
| Approach: | They propose a method that allows teachers to control their distillability via reinforcement fine-tuning (RFT) they propose to use tail noise, off-policy instability, and the teacher–student gap to improve KD. |
| Outcome: | The proposed method outperforms SFT and KD baselines and can be used to protect teachers and students from bottlenecks. |
On the Generalization vs Fidelity Paradox in Knowledge Distillation (2025.findings-acl)
Copied to clipboard
| Challenge: | Knowledge distillation (KD) is a key technique for compressing large language models into smaller ones while preserving performance. |
| Approach: | They propose to use knowledge distillation to compress large language models into smaller ones while preserving performance. |
| Outcome: | The proposed technique improves the performance of smaller models by 10% while providing only marginal benefits for larger models. |
Tutoring Helps Students Learn Better: Improving Knowledge Distillation for BERT with Tutor Network (2022.emnlp-main)
Copied to clipboard
| Challenge: | Existing knowledge distillation approaches for language models have overlooked the difficulty of training examples. |
| Approach: | They propose a framework that controls difficulty of training examples during pre-training by a tutor network. |
| Outcome: | The proposed framework outperforms state-of-the-art KD methods with student models on the GLUE benchmark. |